CVPR 2017 Papers

The choice of motion models is vital in applications like image/video stitching and video stabilization. Conventional methods explored different approaches ranging from simple global parametric models to complex per-pixel optical flow. Mesh-based warping methods achieve a good balance between computational complexity and model flexibility. However, they typically require high quality feature correspondences and suffer from mismatches and low-textured image content. In this paper, we propose a mesh-based photometric alignment method that minimizes pixel intensity difference instead of Euclidean distance of known feature correspondences. The proposed method combines the superior performance of dense photometric alignment with the efficiency of mesh-based image warping. It achieves better global alignment quality than the feature-based counterpart in textured images, and more importantly, it is also robust to low-textured image content. Abundant experiments show that our method can handle a variety of images and videos, and outperforms representative state-of-the-art methods in both image stitching and video stabilization tasks.

Convolutional networks for image classification progressively reduce resolution until the image is represented by tiny feature maps in which the spatial structure of the scene is no longer discernible. Such loss of spatial acuity can limit image classification accuracy and complicate the transfer of the model to downstream applications that require detailed scene understanding. These problems can be alleviated by dilation, which increases the resolution of output feature maps without reducing the receptive field of individual neurons. We show that dilated residual networks (DRNs) outperform their non-dilated counterparts in image classification without increasing the model's depth or complexity. We then study gridding artifacts introduced by dilation, develop an approach to removing these artifacts (`degridding'), and show that this further increases the performance of DRNs. In addition, we show that the accuracy advantage of DRNs is further magnified in downstream applications such as object localization and semantic segmentation.

Material recognition for real-world outdoor surfaces has become increasingly important for computer vision to support its operation "in the wild." Computational surface modeling that underlies material recognition has transitioned from reflectance modeling using in-lab controlled radiometric measurements to image-based representations based on internet-mined images of materials captured in the scene. We propose to take a middle-ground approach for material recognition that takes advantage of both rich radiometric cues and flexible image capture. We realize this by developing a framework for differential angular imaging, where small angular variations in image capture provide an enhanced appearance representation and significant recognition improvement. We build a large-scale material database, Ground Terrain in Outdoor Scenes (GTOS) database, geared towards real use for autonomous agents. The database consists of over 30,000 images covering 40 classes of outdoor ground terrain under varying weather and lighting conditions. We develop a novel approach for material recognition called a Differential An- gular Imaging Network (DAIN) to fully leverage this large dataset. With this novel network architecture, we extract characteristics of materials encoded in the angular and spatial gradients of their appearance. Our results show that DAIN achieves recognition performance that surpasses single view or coarsely quantized multiview images. These results demonstrate the effectiveness of differential angular imaging as a means for flexible, in-place material recognition.

Relationships among objects play a crucial role in image understanding. Despite the great success of deep learning techniques in recognizing individual objects, reasoning about the relationships among objects remains a challenging task. Previous methods often treat this as a classification problem, considering each type of relationship (e.g. "ride") or each distinct visual phrase (e.g. "person- ride-horse") as a category. Such approaches are faced with significant difficulties caused by the high diversity of visual appearance for each kind of relationships or the large number of distinct visual phrases. We propose an integrated framework to tackle this problem. At the heart of this framework is the Deep Relational Network, a novel formulation designed specifically for exploiting the statistical dependencies between objects and their relationships. On two large data sets, the proposed method achieves substantial improvement over state-of-the-art.

Most state-of-the-art text detection methods are specific to horizontal Latin text and are not fast enough for real-time applications. We introduce Segment Linking (SegLink), an oriented text detection method. The main idea is to decompose text into two locally detectable elements, namely segments and links. A segment is an oriented box covering a part of a word or text line; A link connects two adjacent segments, indicating that they belong to the same word or text line. Both elements are detected densely at multiple scales by an end-to-end trained, fully-convolutional neural network. Final detections are produced by combining segments connected by links. Compared with previous methods, SegLink improves along the dimensions of accuracy, speed, and ease of training. It achieves an f-measure of 75.0% on the standard ICDAR 2015 Incidental (Challenge 4) benchmark, outperforming the previous best by a large margin. It runs at over 20 FPS on 512x512 images. Moreover, without modification, SegLink is able to detect long lines of non-Latin text, such as Chinese.

Detecting masked faces (i.e., faces with occlusions) is a challenging task due to two main reasons: 1)the absence of large datasets of masked faces, and 2)the absence of facial cues from the masked regions. To address these issues, this paper first introduces a dataset with 30,811 Internet images and 35,806 annotated MAsked FAces, which is denoted as MAFA. Different from many previous datasets, each annotated face in MAFA is partially occluded by mask. By analyzing the characteristics of masked faces, we propose LLE-CNNs that detect masked face via three major modules. The proposal module first combines two pre-trained CNNs to extract candidate facial regions from the input image and represent them with high dimensional descriptors. After that, the embedding module turns such descriptors into vectors of weights with respect to the components in pre-trained dictionaries of representative normal faces and non-faces by using locally linear embedding. In this manner, missing facial cues in the masked regions can be largely recovered, and the influences of noisy cues introduced by diversified masks can be greatly alleviated. Finally, the verification module takes the weight vectors as input and identifies real facial regions as well as their accurate positions by jointly performing the classification and regression tasks within unified CNNs. Experimental results on MAFA show that the proposed approach significantly outperforms 6 state-of-the-arts by at least 15.6% in detecting masked faces.

Pixel wise image labeling is an interesting and challenging problem with great significance in the computer vision community. In order for a dense labeling algorithm to be able to achieve accurate and precise results, it has to consider the dependencies that exist in the joint space of both the input and the output variables. An implicit approach for modeling those dependencies is by training a deep neural network that, given as input an initial estimate of the output labels and the input image, it will be able to predict a new refined estimate for the labels. In this context, our work is concerned with what is the optimal architecture for performing the label improvement task. We argue that the prior approaches of either directly predicting new label estimates or predicting residual corrections w.r.t. the initial labels with feed-forward deep network architectures are sub-optimal. Instead, we propose a generic architecture that decomposes the label improvement task to three steps: 1) detecting the initial label estimates that are incorrect, 2) replacing the incorrect labels with new ones, and finally 3) refining the renewed labels by predicting residual corrections w.r.t. them. Furthermore, we explore and compare various other alternative architectures that consist of the aforementioned Detection, Replace, and Refine components. We extensively evaluate the examined architectures in the challenging task of dense disparity estimation (stereo matching) and we report both quantitative and qualitative results on three different datasets. Finally, our dense disparity estimation network that implements the proposed generic architecture, achieves state-of-the-art results in the KITTI 2015 test surpassing prior approaches by a significant margin. We plan to release the Torch code that implements the paper in: https://github.com/gidariss/DRR_struct_pred/ .

Today's person detection methods work best when people are in common upright poses and appear reasonably well spaced out in the image. However, in many real images, that's not what people do. People often appear quite close to each other, e.g., with limbs linked or heads touching, and their poses are often not pedestrian-like. We propose an approach to detangle people in multi-person images. We formulate the task as a region assembly problem. Starting from a large set of overlapping regions from body part semantic segmentation and generic object proposals, our optimization approach reassembles those pieces together into multiple person instances. Since optimal region assembly is a challenging combinatorial problem, we present a Lagrangian relaxation method to accelerate the lower bound estimation, thereby enabling a fast branch and bound solution for the global optimum. As output, our method produces a pixel-level map indicating both 1) the body part labels (arm, leg, torso, and head), and 2) which parts belong to which individual person. Our results on challenging datasets show our method is robust to clutter, occlusion, and complex poses. It outperforms a variety of competing methods, including existing detector CRF methods and region CNN approaches. In addition, we demonstrate its impact on a proxemics recognition task, which demands a precise representation of "whose body part is where" in crowded images.

We address the problem of estimating human pose and body shape from 3D scans over time. Reliable estimation of 3D body shape is necessary for many applications including virtual try-on, health monitoring, and avatar creation for virtual reality. Scanning bodies in minimal clothing, however, presents a practical barrier to these applications. We address this problem by estimating body shape under clothing from a sequence of 3D scans. Previous methods that have exploited body models produce smooth shapes lacking personalized details. We contribute a new approach to recover a personalized shape of the person. The estimated shape deviates from a parametric model to fit the 3D scans. We demonstrate the method using high quality 4D data as well as sequences of visual hulls extracted from multi-view images. We also make available BUFF, a new 4D dataset that enables quantitative evaluation (http://buff.is.tue.mpg.de). Our method outperforms the state of the art in both pose estimation and shape estimation, qualitatively and quantitatively.

There are many scientific, medical and industrial imaging applications where users have full control of the scene illumination and color reproduction is not the primary objective For example, it is possible to co-design sensors and spectral illumination in order to classify and detect changes in biological tissues, organic and inorganic materials, and object surface properties. In this paper, we propose two different approaches to illuminant spectrum selection for surface classification. In the supervised framework we formulate a biconvex optimization problem where we alternate between optimizing support vector classifier weights and optimal illuminants. We also describe a sparse Principal Component Analysis (PCA) dimensionality reduction approach that can be used with unlabeled data. We efficiently solve the non-convex PCA problem using a convex relaxation and Alternating Direction Method of Multipliers (ADMM). We compare the classification accuracy of a monochrome imaging sensor with optimized illuminants to the classification accuracy of conventional RGB cameras with natural broadband illumination.

Deep convolutional neural networks (CNNs) are indispensable to state-of-the-art computer vision algorithms. However, they are still rarely deployed on battery-powered mobile devices, such as smartphones and wearable gadgets, where vision algorithms can enable many revolutionary real-world applications. The key limiting factor is the high energy consumption of CNN processing due to its high computational complexity. While there are many previous efforts that try to reduce the CNN model size or the amount of computation, we find that they do not necessarily result in lower energy consumption. Therefore, these targets do not serve as a good metric for energy cost estimation. To close the gap between CNN design and energy consumption optimization, we propose an energy-aware pruning algorithm for CNNs that directly uses the energy consumption of a CNN to guide the pruning process. The energy estimation methodology uses parameters extrapolated from actual hardware measurements. The proposed layer-by-layer pruning algorithm also prunes more aggressively than previously proposed pruning methods by minimizing the error in the output feature maps instead of the filter weights. For each layer, the weights are first pruned and then locally fine-tuned with a closed-form least-square solution to quickly restore the accuracy. After all layers are pruned, the entire network is globally fine-tuned using back-propagation. With the proposed pruning method, the energy consumption of AlexNet and GoogLeNet is reduced by 3.7x and 1.6x, respectively, with less than 1% top-5 accuracy loss. We also show that reducing the number of target classes in AlexNet greatly decreases the number of weights, but has a limited impact on energy consumption.

We present a novel, purely affinity-based natural image matting algorithm. Our method relies on carefully defined pixel-to-pixel connections that enable effective use of information available in the image and the trimap. We control the information flow from the known-opacity regions into the unknown region, as well as within the unknown region itself, by utilizing multiple definitions of pixel affinities. This way we achieve significant improvements on matte quality near challenging regions of the foreground object. Among other forms of information flow, we introduce color-mixture flow, which builds upon local linear embedding and effectively encapsulates the relation between different pixel opacities. Our resulting novel linear system formulation can be solved in closed-form and is robust against several fundamental challenges in natural matting such as holes and remote intricate structures. While our method is primarily designed as a standalone natural matting tool, we show that it can also be used for regularizing mattes obtained by various sampling-based methods. Our evaluation using the public alpha matting benchmark suggests a significant performance improvement over the state-of-the-art.

Shadow removal is a challenging task as it requires the detection/annotation of shadows as well as semantic understanding of the scene. In this paper, we propose an automatic and end-to-end deep neural network (DeshadowNet) to tackle these problems in a unified manner. DeshadowNet is designed with a multi-context architecture, where the output shadow matte is predicted by embedding information from three different perspectives. The first global network extracts shadow features from a global view. Two levels of features are derived from the global network and transferred to two parallel networks. While one extracts the appearance of the input image, the other one involves semantic understanding for final prediction. These two complementary networks generate multi-context features to obtain the shadow matte with fine local details. To evaluate the performance of the proposed method, we construct the first large scale benchmark with 3088 image pairs. Extensive experiments on two publicly available benchmarks and our large-scale benchmark show that the proposed method performs favorably against several state-of-the-art methods.

We consider the problem of two-frame depth from defocus in conditions unsuitable for existing methods yet typical of everyday photography: a handheld cellphone camera, a small aperture, a non-stationary scene and sparse surface texture. Our approach combines a global analysis of image content---3D surfaces, deformations, figure-ground relations, textures---with local estimation of joint depth-flow likelihoods in tiny patches. To enable local estimation we (1) derive novel defocus-equalization filters that induce brightness constancy across frames and (2) impose a tight upper bound on defocus blur---just three pixels in radius---through an appropriate choice of the second frame. For global analysis we use a novel piecewise-spline scene representation that can propagate depth and flow across large irregularly-shaped regions. Our experiments show that this combination preserves sharp boundaries and yields good depth and flow maps in the face of significant noise, uncertainty, non-rigidity, and data sparsity.

Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections--one between each layer and its subsequent layer--our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR- 10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less memory and computation to achieve high performance. Code and pre-trained models are available at https://github.com/liuzhuang13/DenseNet.

In this paper we propose to learn a mapping from image pixels into a dense template grid through a fully convolutional network. We formulate this task as a regression problem and train our network by leveraging upon manually annotated facial landmarks 'in-the-wild'. We use such landmarks to establish a dense correspondence field between a three-dimensional object template and the input image, which then serves as the ground-truth for training our regression system. We show that we can combine ideas from semantic segmentation with regression networks, yielding a highly-accurate `quantized regression' architecture. Our system, called DenseReg, allows us to estimate dense image-to-template correspondences in a fully convolutional manner. As such our network can provide useful correspondence information as a stand-alone system, while when used as an initialization for Statistical Deformable Models we obtain landmark localization results that largely outperform the current state-of-the-art on the challenging 300W benchmark. We thoroughly evaluate our method on a host of facial analysis tasks, and demonstrate its use for other correspondence estimation tasks, such as the human body and the human ear. DenseReg code is made available at http://alpguler.com/DenseReg.html along with supplementary materials.

Dense captioning is a newly emerging computer vision topic for understanding images with dense language descriptions. The goal is to densely detect visual concepts (e.g., objects, object parts, and interactions between them) from images, labeling each with a short descriptive phrase. We identify two key challenges of dense captioning that need to be properly addressed when tackling the problem. First, dense visual concept annotations in each image are associated with highly overlapping target regions, making accurate localization of each visual concept challenging. Second, the large amount of visual concepts makes it hard to recognize each of them by appearance alone. We propose a new model pipeline based on two novel ideas, joint inference and context fusion, to alleviate these two challenges. We design our model architecture in a methodical manner and thoroughly evaluate the variations in architecture. Our final model, compact and efficient, achieves state-of-the-art accuracy on Visual Genome for dense captioning with a relative gain of 73% compared to the previous best algorithm. Qualitative experiments also reveal the semantic capabilities of our model in dense captioning.

Recent progress on saliency detection is substantial, benefiting mostly from the explosive development of Convolutional Neural Networks (CNNs). Semantic segmentation and saliency detection algorithms developed lately have been mostly based on Fully Convolutional Neural Networks (FCNs). There is still a large room for improvement over the generic FCN models that do not explicitly deal with the scale-space problem. Holisitcally-Nested Edge Detector (HED) provides a skip-layer structure with deep supervision for edge and boundary detection, but the performance gain of HED on saliency detection is not obvious. In this paper, we propose a new saliency method by introducing short connections to the skip-layer structures within the HED architecture. Our framework provides rich multi-scale feature maps at each layer, a property that is critically needed to perform segment detection. Our method produces state-of-the-art results on 5 widely tested salient object detection benchmarks, with advantages in terms of efficiency (0.08 seconds per image), effectiveness, and simplicity over the existing algorithms.

Regularization-based image restoration has remained an active research topic in image processing and computer vision. It often leverages a guidance signal captured in different fields as an additional cue. In this work, we present a general framework for image restoration, called deeply aggregated alternating minimization (DeepAM). We propose to train deep neural network to advance two of the steps in the conventional AM algorithm: proximal mapping and b-continuation. Both steps are learned from a large dataset in an end-to-end manner. The proposed framework enables the convolutional neural networks (CNNs) to operate as a regularizer in the AM algorithm. We show that our learned regularizer via deep aggregation outperforms the recent data-driven approaches as well as the nonlocal-based methods. The flexibility and effectiveness of our framework are demonstrated in several restoration tasks, including single image denoising, RGB-NIR restoration, and depth super-resolution.